Deep Research with MCP 🔬
A powerful, self-contained deep research skill that produces thorough, cited reports from multiple web sources. Uses MCP-configured search tools (minimax websearch or zai-mcp-server websearch_prime).
Prerequisites
Before using this skill, ensure you have configured at least one MCP search tool:
- 1. MiniMax web_search (recommended)
- Configure via
minimax MCP server with
web_search tool
- 2. web-search-prime websearchprime
- Configure via
web-search-prime MCP server with
web_search_prime tool
If neither MCP is configured, the skill will not work.
How It Works
When the user asks for research on any topic, follow this workflow:
Step 0: Check MCP Search Availability
First verify MCP search tool availability:
CODEBLOCK0
Priority:
- 1. INLINECODE4
mcporter call MiniMax.web_search query:"keyword"
- 2. INLINECODE5
CODEBLOCK2
Step 1: Understand the Goal (30 seconds)
Ask 1-2 quick clarifying questions:
- - "What's your goal — learning, making a decision, or writing something?"
- "Any specific angle or depth you want?"
If the user says "just research it" — skip ahead with reasonable defaults.
Step 2: Plan the Research (think before searching)
Break the topic into 3-5 research sub-questions. For example:
- - Topic: "Impact of AI on healthcare"
- What are the main AI applications in healthcare today?
- What clinical outcomes have been measured?
- What are the regulatory challenges?
- What companies are leading this space?
- What's the market size and growth trajectory?
Step 3: Execute Multi-Source Search
Use MCP search tools to query:
CODEBLOCK3
Search strategy:
- - Use 2-3 different keyword variations per sub-question
- Aim for 15-30 unique sources total
- Prioritize: academic, official, reputable news > blogs > forums
Step 4: Deep-Read Key Sources
For the most promising URLs, fetch full content:
CODEBLOCK4
Read 3-5 key sources in full for depth. Don't just rely on search snippets.
Step 5: Synthesize & Write Report
Structure the report as:
CODEBLOCK5
Step 6: Save & Deliver
Save report to the current agent's working directory (auto-resolves to actual runtime path):
CODEBLOCK6
Then deliver:
- - Short topics: Post the full report in chat
- Long reports: Post the executive summary + key takeaways, offer full report as file
Quality Rules
- 1. Every claim needs a source. No unsourced assertions.
- Cross-reference. If only one source says it, flag it as unverified.
- Recency matters. Prefer sources from the last 12 months.
- Acknowledge gaps. If you couldn't find good info on a sub-question, say so.
- No hallucination. If you don't know, say "insufficient data found."
Examples
CODEBLOCK7
For Sub-Agent Usage
When spawning as a sub-agent, include the full research request and context:
CODEBLOCK8
Requirements
- - MCP Search Tool (one of the following):
- Minimax MCP server with
web_search tool (recommended)
- zai-mcp-server with
web_search_prime tool
- - curl (for fetching full pages)
- Configure MCP in your OpenClaw config before using this skill
基于MCP的深度研究 🔬
一项功能强大、自包含的深度研究技能,可从多个网络来源生成详尽且带有引用的报告。使用MCP配置的搜索工具(minimax websearch 或 zai-mcp-server websearch_prime)。
前置条件
使用此技能前,请确保已配置至少一个MCP搜索工具:
- 1. MiniMax web_search(推荐)
- 通过 minimax MCP服务器配置 web_search 工具
- 2. web-search-prime websearchprime
- 通过 web-search-prime MCP服务器配置 web
searchprime 工具
如果未配置任何MCP,该技能将无法工作。
工作原理
当用户请求对任何主题进行研究时,请遵循以下工作流程:
步骤0:检查MCP搜索可用性
首先验证MCP搜索工具的可用性:
bash
通过mcporter检查可用的MCP服务器
mcporter list
优先级:
- 1. MiniMax.web_search
bash
mcporter call MiniMax.web_search query:关键词
- 2. web-search-prime.websearchprime
bash
mcporter call web-search-prime.web
searchprime search_query:关键词
步骤1:理解目标(30秒)
提出1-2个快速澄清问题:
- - 您的目标是什么——学习、做决策,还是撰写内容?
- 是否有特定的角度或深度要求?
如果用户说直接研究就好——则使用合理的默认设置跳过此步骤。
步骤2:规划研究(搜索前思考)
将主题分解为3-5个研究子问题。例如:
- 当前人工智能在医疗保健领域的主要应用有哪些?
- 已测量的临床结果有哪些?
- 面临哪些监管挑战?
- 哪些公司处于该领域的领先地位?
- 市场规模和增长轨迹如何?
步骤3:执行多源搜索
使用MCP搜索工具进行查询:
bash
使用zai-mcp-server websearchprime
mcporter call web-search-prime.web
searchprime search_query:<子问题>
使用minimax web_search
mcporter call MiniMax.web_search query:<子问题>
搜索策略:
- - 每个子问题使用2-3种不同的关键词变体
- 目标总计15-30个独特来源
- 优先级:学术、官方、知名新闻 > 博客 > 论坛
步骤4:深度阅读关键来源
对于最有希望的URL,获取完整内容:
bash
curl -sL | python3 -c
import sys, re
html = sys.stdin.read()
去除标签,提取文本
text = re.sub(<[^>]+>, , html)
text = re.sub(r\s+, , text).strip()
print(text[:5000])
完整阅读3-5个关键来源以获取深度。不要仅依赖搜索片段。
步骤5:综合并撰写报告
按以下结构组织报告:
markdown
[主题]:深度研究报告
生成日期:[日期] | 来源数:[N] | 可信度:[高/中/低]
执行摘要
[3-5句关键发现概述]
1. [第一个主要主题]
[带内联引用的发现]
2. [第二个主要主题]
...
3. [第三个主要主题]
...
关键要点
- - [可操作的见解1]
- [可操作的见解2]
- [可操作的见解3]
来源
- 1. 标题 — [一行摘要]
- ...
方法论
在网页和新闻中搜索了[N]个查询。分析了[M]个来源。
调查的子问题:[列表]
步骤6:保存并交付
将报告保存到当前代理的工作目录(自动解析为实际运行时路径):
bash
mkdir -p research/[slug]
将报告写入 research/[slug]/report.md
然后交付:
- - 简短主题:在聊天中发布完整报告
- 长篇报告:发布执行摘要+关键要点,提供完整报告作为文件
质量标准
- 1. 每个主张都需要来源。 不允许无来源的断言。
- 交叉引用。 如果只有一个来源提及,标记为未验证。
- 时效性很重要。 优先选择过去12个月内的来源。
- 承认差距。 如果无法找到某个子问题的良好信息,请说明。
- 禁止幻觉。 如果不知道,请说未找到足够数据。
示例
研究核聚变能源的当前状态
深入比较2026年Rust与Go在后端服务中的表现
研究引导SaaS业务的最佳策略
美国房地产市场目前的情况如何?
子代理使用
当作为子代理生成时,包含完整的研究请求和上下文:
sessions_spawn(
task: 对[主题]进行深度研究。遵循deep-research-with-mcp SKILL.md工作流程。
目标:[用户的目标]
特定角度:[任何具体内容]
将报告保存到research/[slug]/report.md(相对于代理的工作目录)
完成后,用关键发现唤醒主会话。,
label: research-[slug],
model: opus
)
要求
- 带有 web_search 工具的Minimax MCP服务器(推荐)
- 带有 web
searchprime 工具的zai-mcp-server
- - curl(用于获取完整页面)
- 使用此技能前,在OpenClaw配置中配置MCP